(M)LLM-powered computer use agents (CUA) are emerging as a transformative technique to automate human-computer interaction. However, existing CUA benchmarks predominantly target GUI agents, whose evaluation methods are susceptible to UI changes and ignore function interactions exposed by application APIs, e.g., Model Context Protocol (MCP). To this end, we propose MCPWorld, the first automatic CUA testbed for API, GUI, and API-GUI hybrid agents. A key principle of MCPWorld is the use of "white-box apps", i.e., those with source code availability and can be revised/re-compiled as needed (e.g., adding MCP support), with two notable advantages: (1) It greatly broadens the design space of CUA, such as what and how the app features to be exposed/extracted as CUA-callable APIs. (2) It allows MCPWorld to programmatically verify task completion by directly monitoring application behavior through techniques like dynamic code instrumentation, offering robust, accurate CUA evaluation decoupled from specific agent implementations or UI states. Currently, MCPWorld includes 201 well curated and annotated user tasks, covering diversified use cases and difficulty levels. MCPWorld is also fully containerized with GPU acceleration support for flexible adoption on different OS/hardware environments. Our preliminary experiments, using a representative LLM-powered CUA framework, achieve 75.12% task completion accuracy, simultaneously providing initial evidence on the practical effectiveness of agent automation leveraging MCP. Overall, we anticipate MCPWorld to facilitate and standardize the benchmarking of next-generation computer use agents that can leverage rich external tools. Our code and dataset are publicly available at https://github.com/SAAgent/MCPWorld.

We present GaRAGe, a large RAG benchmark with human-curated long-form answers and annotations of each grounding passage, allowing a fine-grained evaluation of whether LLMs can identify relevant grounding when generating RAG answers. Our benchmark contains 2366 questions of diverse complexity, dynamism, and topics, and includes over 35K annotated passages retrieved from both private document sets and the Web, to reflect real-world RAG use cases. This makes it an ideal test bed to evaluate an LLM's ability to identify only the relevant information necessary to compose a response, or provide a deflective response when there is insufficient information. Evaluations of multiple state-of-the-art LLMs on GaRAGe show that the models tend to over-summarise rather than (a) ground their answers strictly on the annotated relevant passages (reaching at most a Relevance-Aware Factuality Score of 60%), or (b) deflect when no relevant grounding is available (reaching at most 31% true positive rate in deflections). The F1 in attribution to relevant sources is at most 58.9%, and we show that performance is particularly reduced when answering time-sensitive questions and when having to draw knowledge from sparser private grounding sources.

The paper addresses two critical challenges in language model (LM) evaluation: creating reliable domain-specific benchmarks and understanding knowledge representation during domain adaptation. We introduce a deterministic pipeline that converts raw domain corpora into completion-type benchmarks without relying on LMs or human curation, eliminating benchmark contamination issues while enabling evaluation on the latest domain data. Our approach generates domain-specific keywords and related word lists using TF and Term TF-IDF methods and constructs prompt-target pairs. We evaluate models by measuring their ability to complete these prompts with the correct domain-specific targets, providing a direct assessment of domain knowledge with low computational cost. Through comprehensive experiments across multiple models (GPT-2 medium/XL, Llama-2/3.1, OLMo-2, Qwen-2, Mistral) and domains, we demonstrate that our benchmark strongly correlates with expert-generated benchmarks while providing a more accurate measure of domain knowledge than traditional perplexity metrics. We reveal that domain adaptation happens rapidly in smaller models (within 500 steps) and illustrate a new approach to domain knowledge evaluation in base models during training for early stopping. By extending mechanistic analysis to domain adaptation, we discover that initial-to-mid layers are primarily responsible for attribute extraction, while later layers focus on next token prediction. Furthermore, we show that during adaptation, forgetting begins in the middle layers, where attribute extraction happens and is amplified in later layers. Our work provides both a practical evaluation methodology for domain-specific LMs and novel insights into knowledge representation during adaptation, with implications for more efficient fine-tuning strategies and targeted approaches to mitigate catastrophic forgetting.

Large Vision-Language Models (VLMs) now generate highly detailed, paragraphlength image captions, yet evaluating their factual accuracy remains challenging. Current methods often miss fine-grained errors, being designed for shorter texts or lacking datasets with verified inaccuracies. We introduce DOCCI-Critique, a benchmark with 1,400 VLM-generated paragraph captions (100 images, 14 VLMs) featuring over 10,216 sentence-level human annotations of factual correctness and explanatory rationales for errors, all within paragraph context. Building on this, we develop VNLI-Critique, a model for automated sentence-level factuality classification and critique generation. We highlight three key applications: (1) VNLI-Critique demonstrates robust generalization, validated by state-of-the-art performance on the M-HalDetect benchmark and strong results in CHOCOLATE claim verification. (2) The VNLI-Critique driven AutoRater for DOCCI-Critique provides reliable VLM rankings, showing excellent alignment with human factuality judgments (e.g., 0.98 Spearman). (3) An innovative Critic-and-Revise pipeline, where critiques from VNLI-Critique guide LLM-based corrections, achieves substantial improvements in caption factuality (e.g., a 46% gain on DetailCaps-4870). Our work offers a crucial benchmark alongside practical tools, designed to significantly elevate the standards for fine-grained evaluation and foster the improvement of VLM image understanding. Project page: https://google.github.io/unblocking-detail-caption

Large language models (LLMs) hold great promise for educational applications, particularly in intelligent tutoring systems. However, effective tutoring requires alignment with pedagogical strategies - something current LLMs lack without task-specific adaptation. In this work, we explore whether fine-grained annotation of teacher intents can improve the quality of LLM-generated tutoring responses. We focus on MathDial, a dialog dataset for math instruction, and apply an automated annotation framework to re-annotate a portion of the dataset using a detailed taxonomy of eleven pedagogical intents. We then fine-tune an LLM using these new annotations and compare its performance to models trained on the original four-category taxonomy. Both automatic and qualitative evaluations show that the fine-grained model produces more pedagogically aligned and effective responses. Our findings highlight the value of intent specificity for controlled text generation in educational settings, and we release our annotated data and code to facilitate further research.

Large Language Models have shown remarkable capabilities in the NLP domain. Their effectiveness can mainly be attributed to their ability to adapt to an array of downstream tasks. However, generally, full fine-tuning is a computationally expensive job. To mitigate this, many techniques have been developed that prime efficiency, a prominent one being Low-Rank Adaptation (LoRA). However, LoRA and its variants employ re-parametrized additive updates. In this paper, we propose Low-Rank Multiplicative Adaptation (LoRMA), which shifts the paradigm of additive updates to a richer space of matrix multiplicative transformations. We tackle challenges such as computational complexity and rank bottleneck of matrix multiplication by effectively re-ordering operations and introducing rank inflation strategies. We conduct extensive experiments to demonstrate the effectiveness of our approach in terms of various evaluation metrics.

Stance detection identifies the viewpoint expressed in text toward a specific target, such as a political figure. While previous datasets have focused primarily on tweet-level stances from established platforms, user-level stance resources, especially on emerging platforms like Bluesky remain scarce. User-level stance detection provides a more holistic view by considering a user's complete posting history rather than isolated posts. We present the first stance detection dataset for the 2024 U.S. presidential election, collected from Bluesky and centered on Kamala Harris and Donald Trump. The dataset comprises 16,044 user-target stance pairs enriched with engagement metadata, interaction graphs, and user posting histories. PolitiSky24 was created using a carefully evaluated pipeline combining advanced information retrieval and large language models, which generates stance labels with supporting rationales and text spans for transparency. The labeling approach achieves 81\% accuracy with scalable LLMs. This resource addresses gaps in political stance analysis through its timeliness, open-data nature, and user-level perspective. The dataset is available at https://doi.org/10.5281/zenodo.15616911

Despite recent advances in retrieval-augmented generation (RAG) for video understanding, effectively understanding long-form video content remains underexplored due to the vast scale and high complexity of video data. Current RAG approaches typically segment videos into fixed-length chunks, which often disrupts the continuity of contextual information and fails to capture authentic scene boundaries. Inspired by the human ability to naturally organize continuous experiences into coherent scenes, we present SceneRAG, a unified framework that leverages large language models to segment videos into narrative-consistent scenes by processing ASR transcripts alongside temporal metadata. SceneRAG further sharpens these initial boundaries through lightweight heuristics and iterative correction. For each scene, the framework fuses information from both visual and textual modalities to extract entity relations and dynamically builds a knowledge graph, enabling robust multi-hop retrieval and generation that account for long-range dependencies. Experiments on the LongerVideos benchmark, featuring over 134 hours of diverse content, confirm that SceneRAG substantially outperforms prior baselines, achieving a win rate of up to 72.5 percent on generation tasks.

Machine learning is advancing rapidly, with applications bringing notable benefits, such as improvements in translation and code generation. Models like ChatGPT, powered by Large Language Models (LLMs), are increasingly integrated into daily life. However, alongside these benefits, LLMs also introduce social risks. Malicious users can exploit LLMs by submitting harmful prompts, such as requesting instructions for illegal activities. To mitigate this, models often include a security mechanism that automatically rejects such harmful prompts. However, they can be bypassed through LLM jailbreaks. Current jailbreaks often require significant manual effort, high computational costs, or result in excessive model modifications that may degrade regular utility. We introduce TwinBreak, an innovative safety alignment removal method. Building on the idea that the safety mechanism operates like an embedded backdoor, TwinBreak identifies and prunes parameters responsible for this functionality. By focusing on the most relevant model layers, TwinBreak performs fine-grained analysis of parameters essential to model utility and safety. TwinBreak is the first method to analyze intermediate outputs from prompts with high structural and content similarity to isolate safety parameters. We present the TwinPrompt dataset containing 100 such twin prompts. Experiments confirm TwinBreak's effectiveness, achieving 89% to 98% success rates with minimal computational requirements across 16 LLMs from five vendors.

This paper introduces PROTEUS, a fully automated system that produces data-driven hypotheses from raw data files. We apply PROTEUS to clinical proteogenomics, a field where effective downstream data analysis and hypothesis proposal is crucial for producing novel discoveries. PROTEUS uses separate modules to simulate different stages of the scientific process, from open-ended data exploration to specific statistical analysis and hypothesis proposal. It formulates research directions, tools, and results in terms of relationships between biological entities, using unified graph structures to manage complex research processes. We applied PROTEUS to 10 clinical multiomics datasets from published research, arriving at 360 total hypotheses. Results were evaluated through external data validation and automatic open-ended scoring. Through exploratory and iterative research, the system can navigate high-throughput and heterogeneous multiomics data to arrive at hypotheses that balance reliability and novelty. In addition to accelerating multiomic analysis, PROTEUS represents a path towards tailoring general autonomous systems to specialized scientific domains to achieve open-ended hypothesis generation from data.